Replays: one attempt on no real model, say why a step was skipped; archive #340 and #338 #348

Merged
cmoriarty merged 2 commits from claude/replay-fixes-340 into main 2026-10-30 16:19:04 -04:00
Owner

Follow-up to #343, from verifying it on run 159's Visual check (refs #340, #338).

  • A scripted replay is one attempt and never reaches a real model. Replaying run 159 scripted took 6.3 min. The attempt failed once its captured turns ran out, mediation tried it again, and attempt 2 started on Zen's deepseek, a real model running on the laptop. Replays now run with mediation_max 0, and the fake Free profile excludes Zen. Same replay: 24 s.
  • A skipped replay says why. It prints the skip reason, the failed step before it and the end of that step's output. Docker's disk was full, so the dependencies step failed.
  • Free replays prune their previous image. Builds are labelled, and the one a build replaces is removed. A day of 2 GB rebuilds had filled the disk.
  • Interrupting a replay no longer leaves the stack running. The runner stops any stack it never handed over.
  • The summary prints the judge's verdict note.
  • replay-a-step and free-judge-holds-the-browser are archived.

Merging with [skip ci]: nothing here changes what production runs.

🤖 Generated with Claude Code

Follow-up to #343, from verifying it on run 159's Visual check (refs #340, #338). - **A scripted replay is one attempt and never reaches a real model.** Replaying run 159 scripted took 6.3 min. The attempt failed once its captured turns ran out, mediation tried it again, and attempt 2 started on Zen's deepseek, a real model running on the laptop. Replays now run with `mediation_max` 0, and the fake Free profile excludes Zen. Same replay: 24 s. - **A skipped replay says why.** It prints the skip reason, the failed step before it and the end of that step's output. Docker's disk was full, so the dependencies step failed. - **Free replays prune their previous image.** Builds are labelled, and the one a build replaces is removed. A day of 2 GB rebuilds had filled the disk. - **Interrupting a replay no longer leaves the stack running.** The runner stops any stack it never handed over. - The summary prints the judge's verdict note. - `replay-a-step` and `free-judge-holds-the-browser` are archived. Merging with `[skip ci]`: nothing here changes what production runs. 🤖 Generated with [Claude Code](https://claude.com/claude-code)
Replaying run 159's Visual check scripted took 6.3 min: its attempt failed once the captured
turns ran out, mediation tried it again, and attempt 2 started on Zen's deepseek, a real model
with its shell on the laptop. A replay now runs with mediation_max 0, and the fake Free
profile excludes Zen. A replayed step that was skipped prints why and the end of the failed
step's output (a full Docker disk failed the dependencies). A stack the runner never handed
over (an error, a Ctrl-C) is stopped instead of outliving it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Verified by replaying run 159's Visual check (attempt 1) captured from production:
scripted, 24 s, the judge's seven browser subagents refused; on Free, 47 min, the judge
drove the browser itself, no subagents, 6 vision requests (production: 136 in 66 min, no
report), and wrote its report.

The replay summary prints the judge's verdict note. Each Free replay build is labelled, and
the build it replaces is pruned: a day of 2 GB rebuilds had filled Docker's disk.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
cmoriarty deleted branch claude/replay-fixes-340 2026-10-30 16:19:04 -04:00
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
cmoriarty/braid!348
No description provided.